Share on social
Sep 17, 2026
Lorem ipsum
This story appeared in our weekly cybersecurity newsletter Beacon. Sign up to get weekly updates on the latest news, findings and insights, straight to your inbox every Thursday at 10AM.
Top Story This Week:
AI agents being tested by OpenAI involved in cyber-attack on another service, say researchers
Researchers have linked a May campaign targeting RubyGems to AI agents being tested by OpenAI. The campaign saw more than 2,000 package submissions, with researchers identifying several hallmarks of AI-generated activity, including packages carrying “oai” identifiers and behavior they say matched previously observed OpenAI agents. They also identified attempts to use RubyGems and its documentation infrastructure to execute code and access API keys.OpenAI has confirmed that its agents used RubyGems during the period, but said they were using the service for benign tasks and to retrieve public information. RubyGems has also said it cannot determine whether AI agents created or published the packages, although it confirmed the campaign and removed more than 500 malicious packages.
The campaign also shows what AI agents are now capable of: interacting with internet infrastructure at scale and finding and testing opportunities without a human directing every step. (Read more in The Guardian)
Why it Matters
An automated attacker doesn't need to start with a known CVE. It can discover an exposed service, test how it behaves, identify functionality it can abuse and potentially chain multiple opportunities together.
This changes how organizations need to look at their attack surface. Understanding what an automated attacker could discover and exploit from the outside, and how quickly it could do so, is different from simply knowing which vulnerabilities are present.As AI makes automated discovery and exploitation more capable, defenders need to understand their environments from the attacker's perspective, and do so continuously.
For Practitioners
Look beyond CVEs and consider what an automated attacker could discover across exposed services, forgotten assets, misconfigurations and unexpected functionality.
Test the attacker's view by identifying the systems and services an external attacker can reach, and look for ways separate components could be chained together.
Validate what is exploitable. Prioritize action based on what an attacker can actually reach and exploit, rather than just relying on vulnerability severity.
For Security Leaders
AI agents are making automated discovery and exploitation faster and more scalable, shrinking the time defenders have to respond.
If attackers can continuously discover and test your environment, a daily or weekly scan leaves long gaps in your visibility. Continuous discovery and validation means your understanding of the attack surface can keep pace with changes to your environment, and with the increasing speed or automated attacks.
This kind of exposure is not only external: any organization running agents with broad tool or network access is itself a potential source of unplanned exposure. Treat what your own agents can reach and do as part of the attack surface you continuously validate.
Discover More
How AI is Collapsing Exploitation Timelines
AI is changing the economics of vulnerability discovery, exploit development, and attack execution - making all three faster, cheaper, and accessible to a much broader range of threat actors. Read our latest blog to understand how to shift from reacting to exploitation toward preempting it - finding and fixing what matters before attackers get there.
Preemptive Threat Exposure Management in the Age of AI
While frontier AI models are dramatically accelerating cyber defense, using them without proper context often floods security teams with overwhelming noise rather than actionable signal. This blog explains how elite researchers are practically applying AI to find critical zero-days, and why staying ahead of the curve requires an evolution to Preemptive Threat Exposure Management.
Weekly News Digest
Active exploitation of Cisco Secure Firewall Management Center vulnerabilities
Cisco Talos reported three threat clusters exploiting two Secure Firewall Management Centre (FMC) vulnerabilities: CVE-2026-20079 (authentication bypass, CVSS 10.0) and CVE-2026-20316 (static credential vulnerability). Activity included web shells, reverse shells, credential theft, and ransomware deployment. One cluster was attributed to Qilin affiliates (UAT-11988); another used Cyclops Blink, previously associated with Sandworm (UAT-11823). The last observed cluster was UAT-12197.
Detecting and countering misuse of AI: September 2026
In its September 2026 threat intelligence report, Anthropic details several instances of AI-augmented cyber operations. Of particular interest is what Anthropic characterizes as the shift from AI as an assistant to an orchestrator of attacks, which was observed in the majority of operations described in its report. Specific examples include the use of AI to increase operational speed by GTG-20006 - which has been linked to Russian state-nexus actor Midnight Blizzard - and the acceleration of smash-and-grab opportunist attacks by financially-motivated cybercriminal actors such as ShinyHunters.
The pro-Russian hacktivist group NoName057(16) has relaunched #OpJapan, a DDoS campaign targeting Japanese organisations due to Japan’s continued support for Ukraine and NATO. Between the 24th-30th of August, the group claimed 66 attacks against 26 organisations, primarily affecting government portals and maritime/logistics companies, with attacks focusing on resource-intensive web functions such as search, login, and contact forms.
